Upload
others
View
0
Download
0
Embed Size (px)
Citation preview
[i]
Heaven’s Light is Our Guide
DEPARTMENT OF COMPUTER SCIENCE & ENGINEERING
RajshahiUniversity of Engineering & Technology, Bangladesh
Automatic Extraction of Road Network from Very High
Resolution Satellite Imagery Using Classification Technique
Author
Abdullah-Al-Zubaer Imran
Roll:073001
Department of Computer Science & Engineering
Rajshahi University of Engineering & Technology
Supervised by
Boshir Ahmed
Head & Asstt. Professor
Department of Computer Science & Engineering
Rajshahi University of Engineering & Technology
[ii]
ACKNOWLEDGEMENT
First of all, I am grateful to my supervisor Boshir Ahmed for his advice, support, and encouragement over the last year. He is an excellent mentor and every conversation I have had with him has been a pleasure and a learning experience. Besides his priceless academic assistance, he also did all that he could help me and make me confident enough.
I am also like to thank my parents who, from the very first day of my education, have provided constant guidance and encouragement.
Next I would like to thank my friends for their continued support. They provided the full encouragement required throughout the entire journey of this work.
Finally, I like to thank my favorite one, the Department of Computer Science & Engineering and all the stuffs working hard to maintain its importance.
06 September, 2012 RUET, Rajshahi
Abdullah-Al-Zubaer Imran
[iii]
Heaven’s Light is Our Guide
DEPARTMENT OF COMPUTER SCIENCE & ENGINEERING
Rajshahi University of Engineering & Technology, Bangladesh
CERTIFICATE
This is to certify that the thesis on “Automatic Extraction of Road Network from Very
High Resolution Satellite Imagery using classification Technique” by Abdullah-Al-
Zubaer Imran, roll: 073001, has been carried out by my supervision.
------------------------------------------------------------------------
Boshir Ahmed
Head &Asstt. Professor
Department of Computer Science &Engineering
Rajshahi University of Engineering &Technology
Rajshahi-6204
September 06, 2012
RUET, Rajshahi
[iv]
ABSTRACT
This paper presents a simplified classification technique for road network extraction from
very high resolution satellite images. We know that satellite images are huge source for
geo-spatial information. For Geographic Information System (GIS), exact road
information of urban areas is very much necessary.Satellite images provide much of the
information for Geographic Information System (GIS). But in satellite images, there are
multiple layers that contain roads, buildings, trees, vehicles and other high density
objects. So, our goal is to extract only roads from the high resolution satellite imagery.
Automatic extraction of road network is one of the important fields to work with. This
paper proposes the method in which at first, by automatic segmentation all the
components belonging to the road areas are separated from other components belonging
to the other objects in the satellite image. Then with all the components belonging to the
road areas are used to create the final road network.The most important thing that can be
achieved in this way is huge amount of flexibility. Actually four sorts of imageries were
carried out for the implementation purpose. Though the results in terms of accuracy fall
short in comparison with the other available techniques, but have higher flexibility rate
and lower execution rate.
Keywords: Automatic road extraction, high resolution satellite imagery, automatic
segmentation, road component, Geographic Information System (GIS).
[v]
CONTENTS
ACKNOWLEDGEMENT CERTIFICATE ABSTRACT CHAPTER-1
Introduction ...................................................................................................... 1
1.1 Problem Statement ............................................................................................. 2
1.2 Motivation .......................................................................................................... 3
1.3 Objectives ......................................................................................................... 3
1.4 Contribution ........................................................................................................ 4
1.5 High Resolution Imagery for Road Extraction ............................................ 4
1.6 Challenges in Automatic Road Extraction ................................................ 5
1.7 Thesis Outlines ............................................................................................ 7
CHAPTER-2
Background & History................................................................................... 8
Introduction ....................................................................................................... 8
2.1 Brief History ................................................................................................ 8 2.2Background of the Work .............................................................................. 9
2.3Conclusion ................................................................................................... 10
CHAPTER-3
Review of Related Work .............................................................................. 11 Introduction ..................................................................................................... 11
3.1 Basics of Road Network Extraction .......................................................... 11
[vi]
3.2 Automatic Seeding ..................................................................................... 12
3.3 Segmentation & Classification .................................................................. 12
3.4 Multi-Resolution Techniques .................................................................... 14
3.5 Artificial Intelligence Approach ................................................................ 15
3.6 Road Tracking Methods ............................................................................ 16
3.7 Snakes ........................................................................................................ 19
3.8 Mathematical Morphology and Filtrate ................................................... 22
Conclusion ....................................................................................................... 23
CHAPTER-4
Components in Road Extraction ............................................................... 25
Introduction ..................................................................................................... 25
4.1 Edge Detectors ........................................................................................... 25
4.2 Canny Edge Detector ................................................................................. 26
4.3 Image Segmentation .................................................................................. 28
4.4 Classification .............................................................................................. 29
4.4.1Supervised Classification ....................................................................... 30
4.4.2 Maximum likelihood Classification ...................................................... 31
4.4.3 Minimum distance Classification ......................................................... 31
4.4.4 Parallelepiped Classification ................................................................ 31
4.4.5 Unsupervised Classification .................................................................. 32
4.5 Mathematical Morphology & Filtrate ...................................................... 33
4.5.1 Connected Component filter ................................................................ 33
4.5.2 Closing ................................................................................................... 34
4.6 Data ............................................................................................................ 35
4.7 Software ..................................................................................................... 35
4.8 Test System Specification .......................................................................... 36
Conclusion ....................................................................................................... 36
CHAPTER-5
Implementation .............................................................................................. 37
[vii]
Introduction ..................................................................................................... 37
5.1 The System Pipeline ................................................................................... 37
5.2 Input High Resolution Satellite Images .................................................... 39
5.3Grayscaling ................................................................................................. 39
5.4Thresholding ............................................................................................... 40
5.5 Edge Detection ........................................................................................... 41
5.6 Convolution Filtering ................................................................................ 42
5.7 Automatic Segmentation ........................................................................... 42
5.8 Component Classification .......................................................................... 43
5.9 Established Road Network ........................................................................ 44
5.10 Smoothing Filtering ................................................................................. 44
5.11 Morphological Processing........................................................................ 45
5.11.1 Dilation ............................................................................................... 45
5.11.2 Erosion ................................................................................................ 46
5.11.3 Closing ................................................................................................ 46
Conclusion ....................................................................................................... 47
CHAPER-6
Performance Evaluation .............................................................................. 48 6.1 Evaluation Method .................................................................................... 48
6.2 Evaluation Mechanism .............................................................................. 48
CHAPTER-7
Conclusion & Recommendation ................................................................ 50 7.1 Conclusions ................................................................................................ 50 7.2 Recommendations for Future Work ......................................................... 51
REFERENCES............................................................................................... 53
[viii]
LIST OF TABLES
1.1 List of currently operational and future optical VHR satellite systems ......... 13
7.1 Experimental results of road images in terms of connected components ...... 49
LIST OF FIGURES
3.1 Flowchart of object-oriented classification .................................................... 14
4.1 Spectral Reflectance curve of 3 land covers .................................................. 30
5.1 The pipeline to road network extraction ........................................................ 38
5.2 Input high resolution satellite images ........................................................... 39
5.3 Grayscale images .......................................................................................... 40
5.4 Binary image ............................................................................................... 41
5.5 Image after edge detection ........................................................................... 42
5.6 Connected components ................................................................................ 43
5.7 Established road image and its complement image ....................................... 44
5.8 Median filter image...................................................................................... 45
5.9 The resultant images after closing operation ................................................ 46
1
Chapter-1 [INTRODUCTION]
Roads play an important role in the field of transportation and communication. In light of
the pivotal role roads play in our daily lives, information pertaining to the location of
roads becomes very much essential. This information not only allows humans to make
informed decisions regarding their environment in general, but also increases efficiency in
the choice of routes for transporting goods and people.
As per advanced technology and technical advancement, at present road locations and
their information are stored and manipulated digitally within geographic databases. The
digital representation is flexible enough to enable numerous Geographic Information
System (GIS) to use road data. These road data enables GIS applications to facilitate a
variety of services which include satellite navigation, route planning, transportation
system modeling [1], health care accessibility planning [2], land cover classification [3]
and even infrastructure management [4].
Two methods are typically used to obtain road data sets, namely ground surveying and
delineating roads from Remotely Sensed imagery [5]. Ground surveying can be conducted
by using devices such as receivers for the Total Station or the Global Positioning System
(GPS). Delineating roads from remotely sensed imagery is known as road extraction.
Road information from remote sensing images is delineated in three ways: i) manual, ii)
semi-automated, iii) and fully automated. Manual extraction of roads from remotely
sensed imagery is a simple stretching operation. However, the operation is impractically
time consuming when the scenes are very complex.
In addition, not only are such complex maps required for large geographic areas, but also
frequent updating is needed. Semi-automated road extraction requires some manual input
to guide the automatic process. Fully automated road extraction removes the need for
human interaction.
Fully automated road extraction systems comprise a variety of algorithms, which can be
roughly categorized into three levels of processing [6]:
2
o The low-level operations that work with the raw image data.
o The medium-level processes that further refine the information gathered by
the low-level algorithms, and
o The high-level algorithms that produce the final road networks.
The higher-level algorithms exhibit aspects of intelligence in their ability to reason on the
structure and location of road networks in a similar fashion of humans.
The goal of this thesis is threefold: to develop a highly flexible road extraction system, to
test various algorithms within this system and finally to propose an attractive road
extraction system. In addition to these goals, a comprehensive observation of the literature
on road extraction is also conducted.
1.1 Problem Statement
Accurate and up-to-date geospatial information about road networks is of great importance
for effective urban and transportation planning, land development, and urban disaster
management [7].
Road information is normally integrated into a Geographical Information System (GIS)
database for effective management. Due to rapid urban development, the GIS database
needs to be updated with timely and accurate road network information. Using traditional
ground survey techniques to collect road data is out of date as it is labor intensive and
time-consuming, particularly in mapping large urban areas.
However, remote sensing has proven to be a powerful technology for spatial data
collection and change detection. Along with the development of innovative sensors and
platforms, road network spatial information can be acquired from aerial and satellite
imagery, which includes optical imagery, Synthetic Aperture Radar (SAR) imagery, Light
Detection and Ranging (LiDAR) range data, and image sequences taken from land-based
mobile mapping systems with different spatial and spectral resolutions [8].
The preferred technology for road extraction is VHR satellite imagery, as it covers larger
areas than aerial imagery and provides updated information on a regular basis. Moreover,
it is more economic than either terrestrial or aerial imaging technologies. Furthermore,
3
unlike SAR and LIDAR data collection, it allows for the extraction of more detailed and
accurate road information.
1.2 Motivation
Although human posses the ability to extract roads from remotely sensed imagery with
fairly high accuracy, the process remains slow and tedious. With the dawn of information
age, computers provide the ability to automate various functions at high speeds.
Automating the road extraction times and thus reduce the costs of creating new maps.
Road extraction research deals not only with the issues of extracting roads, but also
touches on a number of fundamental issues in Digital Images processing (DIP) and HVS.
Research in this area allow the scientific community to develop a clearer understanding of
the manner in which humans are able to combine their cognitive strengths with their vision
systems. These issues are highly relevant with the search in our way to function.
1.3 Objectives
The objectives of this thesis can be stepped in broad senses:
Firstly, a flexible automatic road network extraction system is developed with the intent of
extracting road networks with a high level of accuracy. This system comprises an image
processing pipeline that includes a number of algorithms, ranging from the lowest to the
highest level of processing.
Secondly, the objective is to develop a clearer understanding of the areas in which road
extraction systems suffer the most. This can be achieved by testing a variety of algorithms
at different stages within the processing pipeline.
And finally, the study proposes an attractive automated road network extraction system
with the intent of producing road data sets of high quality.
1.4 Contribution
The main contributions of this thesis can be summarized as:
A generic road network extraction system is presented. The system is flexible and
various configurations to be tested.
4
Various new approaches are tested within the extraction system. The results are
presented and discussed.
Through the development of this system and the testing of algorithms, a clearer
understanding of road extraction is obtained.
The generic system is fully automated, requiring no human intervention during
processing.
1.5 High Resolution Imagery for Road Extraction
Nowadays, optical imagery, especially VHR satellite imagery, has received considerable
attention because it provides accurate, spatial information. Table 1.1 shows the main 4
parameters of currently operational and future optical VHR satellite systems.
All of the VHR satellites mentioned in Table 1.1 simultaneously collect panchromatic
(Pan) and multispectral (MS) images at higher and lower spatial resolution, respectively.
Most MS images are taken within the visible and near-infrared wavelength (VNIR) and
are recorded into 3 or 4 multispectral bands [9].
The image-fusion techniques have been developed to combine Pan and MS images to
generate a high-resolution, pan-sharpened-color image. To date, the PCI Pansharp module
produces the best Pan-MS fusion results among all commercially available software tools
[10]. By image pan-sharpening, the highest-resolution colored satellite imagery available
is GeoEye-1 imagery, which reaches 0.5m spatial resolution.
Currently, DigitalGlobe is developing a VRH imaging satellite named WorldView with a
higher spatialresolution of 0.25m (Pan), and 1m (MS), respectively. Furthermore, there
will be 8 MS bandsin total within the VNIR range. This occurrence proves that spatial
resolution may be furtherimproved; eventually matching the resolution of aerial imagery
(around 0.15m). It is vital tounderstand aerial orthoimage in the event that spatial
resolution of satellite imagery reachesthe level of aerial photographs.
Thus, the images used in this thesis will include pansharpened color GeoEye-1 images and
aerial orthoimages.
5
Table 1.1 List of currently operational and future optical VHR satellite systems
Optical
Satellite
Spatial
resolution (m)
and (#bands)
PAN VNIR
Swath(km) Repeat
cycle
(days)
Year launch
IKONOS 1 4(4) 11 3 1999
QuickBird 2 0.6 2.5(4) 16 3 2001
Orb View-3 1 4(4) 8 3 2003
KOMSAT-2 1 4(4) 15 28 2004
Resurs DK-1 1 2-3(3) 4.7-28.3 6 2006
WorldView-1 0.55 1(4) 17.6 1.7-5.9 2007
GeoEye-1 0.41 1.65(4) 15.2 1-3 2008
Worldview-2 0.25 1(4) 16.4 1-4 2009
Pleiades-1 and
2
0.7 2.8(4) 20 1-2 2009-2010
Source: (Zlatanova and Li, 2008)
1.6 Challenges in Automatic Road Extraction
Road characteristics in general can be classified as radio-metrical, geometrical,
topological, and contextual characteristics [11]. Together, they correspond with the
challenges listed below:
(1) Radio-metrical characteristics and challenges: The road surface is made from
different materials (e.g., asphalt and cement). On the satellite and aerial imagery,
the road surface displays different grayscales representing the construction phase
of the road.
Additionally, adjacent regions are often different. Sometimes, even building roofs
and parking lots appear to similar spectral information to roads. When multiplied
and inconsistent intensity of roads arise and spectral similarity of neighborhood
6
areas increase, creating radio-metrical models for roads will become more
difficult.
(2) Geometric characteristics and challenges: Since their lengths are far larger than
their widths, roads appear as elongated regions in VHR imagery. Roads vary in
width and curvatures. In each road, the width and direction change smoothly.
Various road shapes may increase the difficulty of building geometric models.
(3) Topological characteristics and challenges: Roads form networks when they link
certain places together. In some areas, roads and surrounding buildings are
connected through driveways and alleys. This may cause confusion between
them on the VHR imagery since they both carry similar physical characteristics.
(4) Context characteristics and challenges: The appearances of vehicles, pedestrian
lines, overpasses and shadows cast by trees and buildings can produce a negative
influence on homogeneous intensity and connectivity of roads. In some cases,
shadows may cover certain sections of a road. On the other hand, pedestrian
crossings and trees located on roadsides may help to imply the presence of roads
on a VHR image.
To build models of roads in a large area is difficult as proven by the characteristics and
challenges listed above. Methods that had already been developed for medium to coarse
imagery may not apply to VRH imagery. This is because there is more noise (e.g. vehicles,
traffic lines, shadows, etc.) to lead to blockage problem in VRH imagery [12].
Furthermore, there is another problem. It is difficult to separate the certain roads from
their surroundings with similar spectral information, leading to leakage problem. Thus, a
new methodology for road extraction in VHR imagery is required.
1.7 Thesis Outlines
The rest of the thesis consists of the following six chapters:
Chapter 2 presents a brief history and background for the appropriateness of the thesis.
Chapter 3 provides a review of previous studies on road extraction from VHR imagery.
7
Chapter 4 presents a description of the road network extraction components
Chapter 5 illustrates the detailed implementation of the proposed road extraction system.
Chapter 6 presents a qualitative and quantitative evaluation method. It also focuses on an
evaluation of the developed system’s performance.
Chapter 7 provides the conclusions based on the findings of this study and
recommendations for future research.
8
Chapter-2
[BACKGROUND & HISTORY]
The process whereby roads are extracted from remotely sensed imagery is rather elaborate
and involves a number of techniques from fields within computer science. The aim of this
chapter is that it provides background information and history regarding the fields
involved in the road network extraction process
Introduction The process whereby roads are extracted from remotely sensed imagery is rather elaborate
and involves a number of techniques from numerous fields within computer science. The
aim of this chapter is therefore twofold: firstly it provides background information
regarding the fields involved in the extraction process; secondly, it offers a review of the
history of road extraction system from very high resolution satellite images.
2.1 Brief History
The digital processing of remotely sensed data, with the intent of retrieving or enhancing
man-made features, dates back to the mid 1970s [13,14]. The field of feature extraction
and specially road network extraction has since advanced significantly.
A 2003 review conducted by Mena [7] cites more than 250 road network extraction
studies, which serves as an indication of the vast amount of working effort already
conducted in this field. Owing to the huge amount of studies involved, Mena provides an
overview of the various extraction methods rather than a detailed description. Mena
organizes the studies into various categories, according to technique and the level of
processing.
The technologies used to create road maps evolved significantly from their humble
origins. The exact dates of the first cartographic maps are not clear, but the early
Babylonian maps date back to approximately the 14th-12th century BC. Subsequent maps
created by the ancient Greeks and Romans date back to the 6th century BC. These early
9
maps were created by combining mathematical techniques with observations made by
explorers.
2.2 Background of the Work
Technological advances have changed the way in which modern cartographic maps are
created. Two methods are currently used to obtain map information[5].
The first one which involves ground surveying where data is collected using traditional
methods, such as total station or through modern approaches, such as mobile mapping
systems. Ground surveying requires a team of surveyors to go in the field physically and
record the location of road structures. The process is time-consuming and costly.
The second approach collects data on road positions by identifying them remotely sensed
imagery. The set-of-the-art production road extraction systems are semi-automated, which
implies that the system is driven by a human operator with elements of automation.
Extracting roads from remotely sensed imagery holds some advantages over ground
surveying. Even though extraction from remotely sensed imagery is still time-consuming,
it is less intensive than ground surveying and has lower skill requirements. Extracting
roads from remotely sensed imagery might require Very High Resolution (VHR) imagery,
which could be quite costly.
Those problems of ground surveying can be summarized as:
The process is time-consuming and costly.
This is a labor-intensive process.
The process is often unproductive.
This is so much obsolete technique.
The surveyors are most often fishy i.e., they look for cash rather than quality work
to provide.
Conclusion
The current operational cartographic road network extraction systems rely on heavily on
human input. Some systems provide a level of automation, but human operators still drive
the process to a large extent. Achieving automation in the field of road network extraction
10
will greatly reduce the time and labor required. The knowledge obtained could also be
made more general and applied to other remotely sensed and even Human Visual System
(HVS) problems. So, fully automatic approach that could be developed be well-suited,
efficient and highly flexible in terms of operability.
11
Chapter 3 [REVIEW OF RELATED WORK]
This chapter reviews the principle of technique for road extraction from high resolution
satellite imagery. Previously performed research methods are also reviewed in this chapter
as well.
Introduction
Field of remote sensing involves huge number of applications; extraction of road network
from very high resolution satellite images is the most important one among them. A very
high resolution image taken by satellite will be provided and after some special
processing, the detected road lines will be produced from that image and finally these lines
must be extracted to establish the required road network. Research on extracting roads
from aerial and satellite imagery started since 1976. Many well-furnished methods have
been proposed during last three decades. Three steps are required for extracting road from
imagery [15]: road finding, road tracking and road linking.
3.1 Basics of Road Network Extraction
The road extraction falls into two categories: automatic and semi-automatic. When human
interaction is required in the first step (road finding), then it is referred to as semi-
automatic approach. When there is no human interaction, then it is referred to as fully
automatic [12]. Since the conventional methods have less appropriateness to be used in the
road extraction approach, therefore the emphasis in this chapter is based on the description
of some methods along with the problem related with them.
In the last three decades, there have been many studies on automatic or semi-automatic
road extraction from aerial and satellite imagery. As a result, many strategies,
methodologies and algorithms for road extraction were presented which had reached
various degree of success [16]. All these studies firstly described the characteristics of the
road reflected from the specific remotely sensed data. After that, a feature model was built
based on the characteristic information. Exploring the more information, the better results
12
can be obtained. According to the applied extraction techniques, the existing road
extraction methods using VHR remotely sensed images can be classified into: multi-
resolution techniques, segmentation and classification methods, artificial intelligence
approaches, snakes, road tracking methods and others [7].
3.2 Automatic Seeding
In terms of road extraction, seeding is the process whereby a marker is placed at certain
points of interest within a road network. These points of interest can include markers along
the centre of the road, points of high curvature or intersections. The seeds are typically
single points but can also be centerline segments or road regions. Though seeding is not an
extraction technique itself, but the markers are used as the initialization points for
extraction techniques such as road trackers and snakes. The seeds can also be used to
generate road models or pattern classes, which can be used to train classifiers in order to
detect road objects in imagery. In addition, seeds can be used in road network construction
algorithms to connect the points using a high-level technique.
Automatic seeding can be categorized as low to medium-level processing techniques, as
they typically receive raw image data or output from a low-level algorithm as input. One
of the most popular approaches to automatic seeding is the detection of parallel edges in
medium to high resolution imagery. Seed points or lines are formed by calculating the
midpoint between the parallel lines, whereas seed segments are typically created by
connecting the end points of the parallel lines to form a rectangle.
3.3 Segmentation & Classification
Segmentation is the automated process of partitioning an image into several clusters that
are homogeneous with respect to some characteristics such as color, texture, reflection
signal, or context, etc [17]. Classification is the process of assigning segmented individual
pixels or homogeneous clusters to specific and more meaningful information classes [18].
These two techniques are used for road extraction in order to obtain a binary image where
the road network is depicted. The aim of segmentation is to add structure to the data,
which allows for faster and more accurate analysis. There are two kinds of classification
procedures: supervised and unsupervised classifications. Unsupervised classification does
not require human interpolation, while the supervised methods need foreknowledge.
13
Image segmentation and classification are used for road extraction in order to obtain a
binary image where the road network is depicted. Researchers use these techniques
frequently to extract road networks from remote sensing images.
Most traditional classification algorithms, such as ISODATA, perform classification using
spectral analysis based on single pixels. The feature extraction techniques utilized by the
eCognition® software consider image objects rather than single pixels when performing
classifications [19]. Figure 3.1 shows the framework of the object-oriented classification
approach.
Figure 3.1: Flowchart of object-oriented classification.
A multi-resolution segmentation is first used to produce image objects, so as to calculate
the features which could be used to classify these objects to a specific class. Another
segmentation and classification will be operated until the best result could be obtained. In
Input: Image
Multi-resolution Segmentation
Image Object
Classification
Output
Acceptable?
14
these processes, the objects produced by segmentation will be propitious to subsequent
classification, and the classified objects will also advantage the subsequent segmentation.
Accordingly, this method is appropriate for the extraction of roads from high resolution
imagery. All the studies mentioned above have the same steps as common grounds, which
are to segment the image, to build a model to define a requirement based on the road
characteristics, and to classify the road class by clusters which satisfy the specific
requirement. However, the segmentation algorithms need to scan the whole image, and
every pixel needs to be proceeded at least once, costing much more time.
Additionally, only when the road areas have a homogeneous color, the segmentation can
generate satisfied segments for subsequent classification. If the road does not have obvious
different color from its neighbor area, the satisfied segments cannot be obtained, which
makes the subsequent classification step difficult to generate an appropriate criterion to
classify.
3.4 Multi-Resolution Techniques
As the objects are represented with more details in VHR imagery, the task of road finding
requires complicated methods for the abstraction of these details. In contrast, roads in
medium and coarse resolution imagery can be extracted by using simple road models. Line
extraction at low resolution can guide detailed image analysis at high resolution and
restrict it to image parts with high probability of containing a road. Many multi-resolution
approaches first generate lower resolution imagery by degrading a high resolution image.
Shneier successively applied a 2*2 filter over the image to replace the four-pixel
neighborhood with the median value, creating a pyramid of images with progressively
lower resolution from which lines was extracted by a line detector for identifying roads in
the higher resolution imagery [20]. Each line in the lower resolution image was assumed
to correspond to an elongated region in the original image, and could be utilized to
identify the position and the extent of these regions. As an alternative, Couloigner and
Ranchin used a wavelet transform to generate pyramid layers [21].
15
Instead of degrading high resolution imagery for multi-resolution analysis, some
researchers integrate different types of satellite imagery for extraction. Bonnefon used
SPOT imagery to approximately identify linear features which are used to identify roads in
IKONOS imagery [22].
Trinder and Wang found distinct advantages by combining the abstraction of the coarser
scale with the detailed information found at the finer scale [15]. Pairs of edges in high
resolution imagery were first identified and then combined with lines in lower resolution
(re-sampled) imagery in order to fully extract the road network.
Baumgartner combined roadsides extracted from the original high resolution image (0.2–
0.5m) and lines extracted from an image of reduced resolution to build the hierarchical
road structure[23]. The road area can be identified if a pair of edges are approximately
parallel, have an approximately homogeneous region between them, and have a
corresponding line in the reduced resolution image.
Road detection in low resolution is the key step for guiding road extraction in high
resolution. Multi-resolution road extraction is appropriate for rural area where the majority
of roads are easily extracted from low resolution image but not for high density downtown
areas where extra edges which represent roads are extracted in low resolution imagery.
3.5 Artificial Intelligence Approach
Artificial intelligence approaches convert human knowledge into an exploitable form,
aimingto carry out reasoning in the same way as a human being [24]. In this way,
combining rules from a variety of sources to build a road model allows the computer to
provide correct, flexible and effective results. Fuzzy logic, neural networks and genetic
algorithms are the most popular mathematical tools used in artificial intelligence systems
[25].
Mohammadzadeh proposed a fuzzy logic approach to detect main road centerline from
pan-sharpened IKONOS images. Depending on the complexity of the scenes, one or up to
three pixels might be manually selected from the road surface and the RGB grey levels of
these pixels were considered as initial values [25]. The best mean values in each band
from fuzzification processing were used in defuzzification to generate a segmented image.
16
After that, morphological algorithm was applied to the segmented image for small paths
removal, holes filing, and centerline extraction. However, road covered by large shadows
cannot be extracted in this method. Moreover, this fuzzy model is mainly based on
spectral information, which may be ascribed to failure for complicated scenes.
Mokhtarzade and Zoej used artificial neural networks for road detection from high
resolution satellite image [26]. The discrimination ability of the network is highly affected
by the choosing of input parameters. After testing multi-spectral IKONOS and Quick-Bird
images, the network can be empowered in road detection by inputting neighbour pixels
and in background detection by inputting the distance of each pixel to the road mean
vector. The successful detection should base on the assumption that roads are
homogeneous area. These two methods make good use of three bands information to
extract road networks. The more characteristics of road are modeled, the more accurate
results will be obtained. However, only spectral and few shape information is used in these
two models, therefore, they are only appropriate for simple scenes where roads are salient.
3.6 Road Tracking Methods
Semi-automatic road tracking by template matching methods seems to be more useful in
operational applications due to the participation of human operators in tracing [23]. This
kind of method is an iterative road segment growing process starting from a set of seed
points automatically or manually selected. The templates can be categorized into two
classes: profile (1-dimension cross-section of the image intensities taken orthogonal to the
direction of a road segment) and rectangular template (2- dimension cross-section of the
image intensities taken orthogonal to the direction of a road segment) based on dimension
of road template. Template matching compares a reference template with the template at a
pixel to predict the road direction.
McKeown and Denlinger introduced a road tracking method based on multiple
cooperative methods which included a surface correlation tracker and a road edge tracker
[27]. When one method failed at one point, the alternative tracking method could be used
to continue the tracking process. However, these two trackers have a lot of limitations,
because they can work only when the road has constant width, gradually or suddenly
changing intensity profile, slowly changing direction and so on.
17
Vosselman and Knecht extracted roads based on least squares matching of grey value
profiles [11]; Kalman filter was also used to continue the process when the profile
matching failed. The two cooperative can trace roads with intersections, flyovers, and
vehicle. This method can settle complex situations such as occlusion by trees, shadows
which are left to the operator.
Baumgartner built a graphical user interface for a profile matching method in a style of
method mentioned in Vosselman and Knecht Therefore, the operator can monitor the
tracking processing, and receive report which describe the problems occurred during the
tracking through this user interface. This tracking tool is fast, but only appropriate for
simple scene such as rural areas.
Shukla applied a path following method to extract road form high-resolution satellite
image by initializing two points to indicate the road direction[28]. Scale space, Edge
detection techniques were used as pre-processing for segmentation and estimation of road
width. The cost minimization technique was used to determine the road direction and
generate next seeds. This method is better than the work of Park and Kim (2001) because
it can generate seeds in different directions at intersection. The limitations are that the
algorithm may not work on the road cast by shadows.
Zhao et al. (2002) imposed a semi-automatic method by matching a rectangular road
template with both road mask and road seeds to extract roads from IKONOS imagery [29].
Road mask is the road pixels generated from maximum likelihood classification, and the
road seeds can be generated by tracing the long edge of the road mask. The problem is that
neither all of extracted road mask are road area nor all of the extracted long edges are road
edge, which would result in misclassification.
Hu et al. (2004) presented a semi-automatic road extraction method based on a piecewise
parabolic model with 0-order continuity, which was constructed by seed points placed by a
human operator [30]. Road extraction became the problem of estimating the unknown
parameters for each piece of parabola, which could be solved by least square template
matching based on the deformable template and the constraint of the geometric model. In
densely populated areas, where roads have sharp turns and orthogonal intersections, a
plenty of seed points are needed to be located, resulting in a degrading the efficiency.
18
Kim [31] used least squares correlation matching to extract road centerlines from
IKONOS images based on the orientation of the initial seed calculated through Burns line
extraction algorithm and a road template built around the seed. The limitations of this
algorithm are that it cannot work on the road with shadows which may terminate the
tracking process, the initial seed must be selected on road central lines by the operator, and
one seed can only extract road with one direction, leading to too many seeds when the
scene is large and complex.
Zhou [32] proposed a framework which consists of the user, a human–computer interface,
computer vision algorithms, knowledge transfer schemes and an evaluation criterion for
semi-automatic road extraction. Extended Kalman filter and particle filter were applied to
solve profile matching issues for road tracking to enhance the robustness of the tracker.
Two profiles were required, one perpendicular to the road direction and the other one
parallel to the road direction.
Another tracking method is based on texture signature. Hu et al. (2007) extracted road
networks from aerial images by tracking road footprints obtained by a spoke wheel
operator based on texture information [30]. The first road footprint was generated from a
road seed initialized by an operator or automatically generated based on rectangular
approximations. Then, a toe-finding algorithm was used to classify footprints for growing
a road tree. Finally, a Bayes decision model based on the area-to-perimeter ratio of the
footprint was applied to prune the paths that leaked into the surroundings. This method can
successfully extract various shapes of roads and intersections, while the footprints often
fail to be generated because of the marking lines or shadows from trees or buildings.
The road tracking methods are mostly semiautomatic at which the running time is
proportional to the area of the road rather than the area of the whole image, and as such
they are all efficient. However, most of the road trackers mentioned above would fail
when they encountered blockage problems caused by radiometric changes due to shadows,
vehicle congestions, pavement changes, lane markings, overpasses, etc [12].
3.7 Snakes
The concept “snake”, also called “active contour model” was first introduced by Kass
which has been used to seek any shape in the image that was smooth and forms a closed
19
contour [33]. Since the boundaries of road network were of diverse shapes including
various degrees of curvature, snakes were well suited for this task. The extraction process
was started by initializing a curve, called snake, close to the object boundary by the
operator. Then, the curve was associated with an objective function which combined
internal smoothness constraints such as bending of a curve with image forces like the
gradient. By optimizing the objective function iteratively, the curve started deforming and
moved towards the desired object boundary. In the end, it completely “shrink-wraped”
around the object [33]
However, traditional snakes are extremely sensitive to parameters and its convergence is
dependent on initial position. Moreover, it has a small capture range since no external
force acts on those points far away from the boundary. Furthermore, it fails to detect
concave boundaries because external force cannot pull control points into boundary
concavity. Since road networks have various degrees of curvature a close initialization
often cannot be provided. As a result, traditional snakes can easily get stuck in an
undesirable local minimum [12].
Neuenschwander proposed the ziplock snake model allowing a user to only specify the
distant end points of the curve, simplifying far less initialization effort. The image
information around the end points was used to provide boundary conditions, Moreover,
when a snake was near its extremities, the image information would be taken into account
at first. Accordingly, this modified snake model yields excellent convergence properties
for the snakes even if the initialization is far away from the solution. They also extended
the snake based approach to ribbon snakes to extract roads in rural area. The extraction
process could be achieved by optimizing the position and the width of the ribbon.
However, it will be stopped in the presence of disturbances[34].
Gruen and Li used either a dynamic programming approach or LSB-Snakes (Least
Squares B-spline Snakes) to extract linear feature after providing a few seed points [35].
Dynamic programming was used to build the cost function among the seeds and solve
optimization problems to extracts roads with the parameter model. LSB-Snakes which
combine least squares template matching [35] and B-spline Snakes [36] improved the
performance of active contour models and controlled blunders such as occlusions very
well.
20
Mayer [37] and Laptev [38] used ribbon snakes to extracted salient roads based on the
detected lines at a coarse scale and the variation of the road width at a fine scale. Non-
salient roads were extracted by connecting two adjacent ends of salient roads with a road
hypothesis, which was then verified based on homogeneity and the constancy of width.
Finally, a closed snake was initialized inside the central area of the junction and expanded
until delineating the junction borders. Nevertheless, cars, traffic islands and road markings
in urban area can block the snake’s movement; therefore, this approach is more intended
for rural areas.
Amo[39] used the region competition algorithm to extract roads from aerial images. The
region competition was a mixed approach which combined region growing techniques
with active contour model. Region growing made the first step faster and region
competition delivers more accurate results. However, this method is only appropriate for
handling roads in agricultural fields, where roads are quite homogeneous and their
homogeneity is sufficiently different from that of their surroundings.
Niu[12] presented a semi-automatic framework for highway extraction based on a
geometric deformable model which referred to the minimization of an objective function
that connects the optimization problem with the propagation of regular curves. After the
seed points were placed at the end of the highway segments, the framework would
incorporate the shape information of a highway segment into the seed point propagation
scheme, thus it successfully prevented the leakage problem and blockage problem.
However, it is based on the assumption that the highway is a continuous ribbon with no
sudden ends or sharp turns, therefore, this method is not appropriate for roads in
residential areas.
Most snake methods are applied on rural area, where roads’ color is homogeneous inside
and different from neighbor areas. In urban areas, various types of features which often
present inside the road areas in VHR imagery, such as cars, traffic islands, road markings,
shadows cast by trees or buildings, which can block the snake’s movement. Furthermore,
expanding snakes can pass over weak junction borders, leading to some leakage problems.
21
3.8 Mathematical Morphology and Filtrate
Mathematical morphology (MM) is a theory and technique for the analysis and
processingof geometrical structures, based on set theory, lattice theory, topology, and
random functions. MM is most commonly applied to digital images, but it can be
employed as well on graphs, surface meshes, solids, and many other spatial structures.
Topological and geometrical continuous-space concepts such as size, shape, convexity,
connectivity, and geodesic distance, were introduced by MM on both continuous and
discrete spaces. MM is also the foundation of morphological image processing, which
consists of a set of operators that transform images according to the above
characterizations.
From mid-1970s to mid-1980s, MM was generalized to grayscale functions and images as
well. Besides extending the main concepts (such as dilation, erosion, etc...) to functions,
this generalization yielded new operators, such as morphological gradients, top-hat
transform and the Watershed (MM's main segmentation approach).
In the 1980s and 1990s, MM gained a wider recognition, as research centers in several
countries began to adopt and investigate the method. MM started to be applied to a large
number of imaging problems and applications.
In 1986, Serra further generalized MM, this time to a theoretical framework based on
complete lattices. This generalization brought flexibility to the theory, enabling its
application to a much larger number of structures, including color images, video, graphs,
meshes, etc. At the same time, Matheron and Serra also formulated a theory for
morphological filtering, based on the new lattice framework.
The 1990s and 2000s also saw further theoretical advancements, including the concepts of
connections and levelings.
In 1993, the first International Symposium on Mathematical Morphology (ISMM) took
place in Barcelona, Spain. Since then, ISMMs are organized every 2–3 years, each time in
a different part of the world: Fontainebleau, France (1994); Atlanta, USA (1996);
Amsterdam, Netherlands (1998); Palo Alto, CA, USA (2000); Sydney, Australia (2002);
22
Paris, France (2005); Rio de Janeiro, Brazil (2007); Groningen, Netherlands (2009); and
Intra (Verbania), Italy (2011).
Conclusion
In conclusion, most methods can extract salient roads in simple scenes such as rural areas.
While for some complicated scenes, various problems may occur in different kinds of
degree, reducing the correctness or completeness of the road extraction.
In multi-resolution, results from the medium and coarse resolution images bring more
effectiveness for the road finding processing in VHR imagery, while results from VHR
imagery provide more detailed information. Nevertheless, this method still cannot detect
roads covered by big shadows, because no clue can be obtained from different resolution
images used for detecting that blockage area.
Segmentation and classification methods usually separate the images into several
homogeneous segments, and propose a rule to detect road area, while the noisy road can
affect the generated segments. Consequently, no satisfied rule can be made to extract all
road areas. Moreover, the running time of segmentation is usually very long, because
every pixel in the whole image needs to be calculated to get a homogeneous segment.
Artificial intelligence approaches usually builds a model, which includes rules with
different weights, the area which satisfies all the rules to some degree can be considered
road area, and nevertheless, rules for very complex scenes are difficult to be exploited into
some exploitable form that can be performed in computers.
Snake methods reduce the road extraction process as the optimization of an objective
function and it can detect the roads with various shape but not the ones with too much
noise.
Road tracking method tracks the road by comparing it with the template to determine the
road direction, thus only pixels on the road area and around it are calculated, however,
blockage on the roads may lead the stop of tracking, and not all detected areas which
matches with the template are guaranteed as the road area.
23
Based on the discussions above, segmentation and classification based automatic
extraction by applying road tracking method into it can be developed without the external
intervention of human.
24
Chapter-4 [COMPONENTS IN ROAD EXTRACTION]
Further from the literature review presented in the previous chapter, a number of methods
are used as autonomous components in the road extraction system. This chapter provides
the available components in the overall extraction process of the road network.
Introduction
Different types of methods for the extraction of road networks have been presented in the
previous chapter. A number of these methods were implemented and used to form part of
a upgraded remote network extraction.
These selected methods are implemented as autonomous components, enabling them to be
stringed together in numerous image processing chain configurations. To observe the
possibility of increasing the accuracy of remote network extraction systems, different
component arrangements are tested.
These specified methods were implemented as modular, point-specific and exchangeable
components. In order to prevent the creation and explosion of illogical connections in the
processing chain, only task-specific components are allowed to be interchangeable. The
generic nature of the components facilitated the necessary flexibility to test a number of
different system configurations.
4.1 Edge Detectors
Edge detection is the approach most frequently for segmenting images based on abrupt
(local) changes in intensity. So, an edge detector works with the goal of detecting the
changes in intensity for the purpose of finding edges that can be accomplished using first-
or second –order derivatives.
According to the intensity profiles there are edge models as:
A step edge involves a transition between two intensity levels occurring ideally over the
distance of 1 pixel.
25
A ramp edge is blurred and noisy, with the degree of blurring determined principally by
limitations in the focusing mechanism and the noise level determined principally by the
electronic components of the imaging system.
A roof edge model is the model of lines through a region with the base (width) of a roof
edge being determined by the thickness and sharpness of the line.
There are three fundamental steps performed in edge detection [40]. The steps are as in the
follow:
1. Image smoothing for noise reduction.
2. Detection edge points- a local operation that extracts from an image all points that are
potential candidates to become edge points.
3. Edge localization- the objective of this step is to select from the candidate edge points
that are true members of the set of points comprising an edge.
As with many other edge detection approaches, the Canny algorithm considers intensity
changes within a local area in digital images. The following section provides a detailed
description of the Canny algorithms.
4.2 Canny Edge Detector
Canny’s approach is based on three basic objectives. These are low error rate, localized
edge points, and single edge point response. Canny seeks to formalize these requirements
and consequently holds the first derivative of a Gaussian function swerves as suitable
approximation for the optimal edge detector. However, the first derivative of the Gaussian
function is implemented by first smoothing the input image with a Gaussian filter, and
subsequently computing the finite difference derivative.
Canny developed an algorithm that describes each of these requirements mathematically.
The algorithm of Canny can be summarized in the following six steps:
1. The first step is to filter out any noise in the original image before trying to locate and
detect any edges. And because the Gaussian filter can be computed using a simple mask, it
is used exclusively in the Canny algorithm. Once a suitable mask has been calculated, the
Gaussian smoothing can be performed using standard convolution methods. A convolution
26
mask is usually much smaller than the actual image. As a result, the mask is slid over the
image, manipulating a square of pixels at a time. The larger the width of the Gaussian
mask, the lower is the detector's sensitivity to noise. The localization error in the detected
edges also increases slightly as the Gaussian width is increased. The Gaussian mask used
in my implementation is shown below.
2.After smoothing the image and eliminating the noise, the next step is to find the edge
strength by taking the gradient of the image. The Sobel operator performs a 2-D spatial
gradient measurement on an image. Then, the approximate absolute gradient magnitude
(edge strength) at each point can be found. The Sobel operator uses a pair of 3x3
convolution masks, one estimating the gradient in the x-direction (columns) and the other
estimating the gradient in the y-direction (rows). They are shown below:
Gx Gy
-1 0 +1
-2 0 +2
-1 0 +1
The magnitude, or EDGE STRENGTH, of the gradient is then approximated using the
formula: |G| = |Gx| + |Gy|
3 . Compute the edge magnitude as G=√Gx2 + Gy
2, and the edge direction as
Ө = tan-1 (Gy/Gx), Gx≠ 0.
4. Discretize the gradient direction Ө of each pixel into one of eight possible neighboring
sectors.
5. After the edge directions are known, non-maximum suppression on the gradient
magnitude G now has to be applied. Non-maximum suppression is used to trace along the
edge in the edge direction and suppress any pixel value (sets it equal to 0) that is not
considered to be an edge. This will give a thin line in the output image.
+1 +2 +1
0 0 0
-1 -2 -1
27
6. Apply hysteresis to the remaining non-maximum gradient magnitudes. The hysteresis
process marks a pixel as an edge pixel if its magnitude value exceed the high threshold
Thigh. The neighbors of the edge pixels are subsequently traced and labeled as edge pixels
if their magnitudes remain above the lower threshold Tlow.
The algorithm output is a bitmap of single pixel-width edges, along with the edge
orientation, Ө, for each of these edge pixels. The Canny edges illustrated in Figure 4.1 b
were created with the following parameter configuration: σ = 1.5, Tlow = 0.0, and Thigh =
0.4.
4.3 Image Segmentation
Various remote sensing feature extraction approaches perform analyses on a per-pixel
basis, where each pixel is analyzed as an individual object in the feature space. The arrival
of very high resolution satellite imagery introduced the option to apply object-based
analysis through segmentation. Object-based analysis considers an entire segment in
feature space, rather than individual pixels. The aim of segmentation is to add structure to
the data, which allows for faster and more accurate analysis [41].
Segmentation algorithms attempt to identify regions with similar characteristics. These
characteristics depend on the problem at hand, but are generally defined by the
discontinuities and similarities of intensity values within an image [40]. Neighboring
pixels with similar intensity values are thus grouped together, while borders are created
where discontinuities occur.
A central problem of segmentation is to distinguish objects from background. For intensity
images (i.e., those represented by point-wise intensity levels) four popular approaches are:
threshold techniques, edge-based methods, region-based techniques, and connectivity-
preserving relaxation methods.
Threshold techniques, which make decisions based on local pixel information, are
effective when the intensity levels of the objects fall squarely outside the range of levels in
the background. Because spatial information is ignored, however, blurred region
boundaries can create havoc.
28
Edge-based methods center around contour detection: their weakness in connecting
together broken contour lines make them, too, prone to failure in the presence of blurring.
A region-based method usually proceeds as follows: the image is partitioned into
connected regions by grouping neighboring pixels of similar intensity levels. Adjacent
regions are then merged under some criterion involving perhaps homogeneity or sharpness
of region boundaries. Over-stringent criteria create fragmentation; lenient ones overlook
blurred boundaries and over-merge. Hybrid techniques using a mix of the methods above
are also popular.
A connectivity-preserving relaxation-based segmentation method, usually referred to as
the active contour model, was proposed recently. The main idea is to start with some
initial boundary shape represented in the form of spline curves, and iteratively modify it
by applying various shrink/expansion operations according to some energy function.
Although the energy-minimizing model is not new, coupling it with the maintenance of an
``elastic'' contour model gives it an interesting new twist. As usual with such methods,
getting trapped into a local minimum is a risk against which one must guard; this is no
easy task.
4.4 Classification
The intent of the classification process is to categorize all pixels in a digital image into one
of several land cover classes, or "themes". This categorized data may then be used to
produce thematic maps of the land cover present in an image. Normally, multispectral data
are used to perform the classification and, indeed, the spectral pattern present within the
data for each pixel is used as the numerical basis for categorization (Lillesand and Kiefer,
1994).
The objective of image classification is to identify and portray, as a unique gray level (or
color), the features occurring in an image in terms of the object or type of land cover these
features actually represent on the ground. Image classification is perhaps the most
important part of digital image analysis. It is very nice to have a "pretty picture" or an
image, showing a magnitude of colors illustrating various features of the underlying
terrain, but it is quite useless unless to know what the colors mean. [42]. Two main
classification methods are Supervised Classification and Unsupervised Classification.
29
Figure 4.1: Spectral Reflectance curve of 3 land covers
4.4.1Supervised Classification
With supervised classification, we identify examples of the Information classes (i.e., land
cover type) of interest in the image. These are called "training sites". The image
processing software system is then used to develop a statistical characterization of the
reflectance for each information class. This stage is often called "signature analysis" and
may involve developing a characterization as simple as the mean or the rage of reflectance
on each bands, or as complex as detailed analyses of the mean, variances and covariance
over all bands. Once a statistical characterization has been achieved for each information
class, the image is then classified by examining the reflectance for each pixel and making
a decision about which of the signatures it resembles most. (Eastman, 1995)
4.4.2 Maximum likelihood Classification
Maximum likelihood Classification is a statistical decision criterion to assist in the
classification of overlapping signatures; pixels are assigned to the class of highest
probability. The maximum likelihood classifier is considered to give more accurateresults
than parallelepiped classification however it is much slower due to extra computations.
We put the word `accurate' in quotes because this assumes that classes in the input data
have a Gaussian distribution and that signatures were well selected; this is not always a
safe assumption.
30
4.4.3 Minimum distance Classification
Minimum distance classifies image data on a database file using a set of 256 possible class
signature segments as specified by signature parameter. Each segment specified in
signature, for example, stores signature data pertaining to a particular class. Only the mean
vector in each class signature segment is used. Other data, such as standard deviations and
covariance matrices, are ignored (though the maximum likelihood classifier uses this).
The result of the classification is a theme map directed to a specified database image
channel. A theme map encodes each class with a unique gray level. The gray-level value
used to encode a class is specified when the class signature is created. If the theme map is
later transferred to the display, then a pseudo-color table should be loaded so that each
class is represented by a different color.
4.4.4 Parallelepiped Classification
The parallelepiped classifier uses the class limits and stored in each class signature to
determine if a given pixel falls within the class or not. The class limits specify the
dimensions (in standard deviation units) of each side of a parallelepiped surrounding the
mean of the class in feature space.
If the pixel falls inside the parallelepiped, it is assigned to the class. However, if the pixel
falls within more than one class, it is put in the overlap class (code 255). If the pixel does
not fall inside any class, it is assigned to the null class (code 0).
The parallelepiped classifier is typically used when speed is required. The draw back is (in
many cases) poor accuracy and a large number of pixels classified as ties (or overlap, class
255).
4.4.5 Unsupervised Classification
Unsupervised classification is a method which examines a large number of unknown
pixels and divides into a number of classed based on natural groupings present in the
image values. unlike supervised classification, unsupervised classification does not require
analyst-specified training data. The basic premise is that values within a given cover type
should be close together in the measurement space (i.e. have similar gray levels), whereas
31
data in different classes should be comparatively well separated (i.e. have very different
gray levels) [42]
The classes that result from unsupervised classification are spectral classed which based
on natural groupings of the image values, the identity of the spectral class will not be
initially known, must compare classified data to some from of reference data (such as
larger scale imagery, maps, or site visits) to determine the identity and informational
values of the spectral classes. Thus, in the supervised approach, to define useful
information categories and then examine their spectral separablity; in the unsupervised
approach the computer determines spectrally separable class, and then define their
information value. [42]
Unsupervised classification is becoming increasingly popular in agencies involved in long
term GIS database maintenance. The reason is that there are now systems that use
clustering procedures that are extremely fast and require little in the nature of operational
parameters. Thus it is becoming possible to train GIS analysis with only a general
familiarity with remote sensing to undertake classifications that meet typical map accuracy
standards. With suitable ground truth accuracy assessment procedures, this tool can
provide a remarkably rapid means of producing quality land cover data on a continuing
basis.
4.5 Mathematical Morphology & Filtrate
The results obtained by road classifiers or the road plausibility values produced by the
prvious operations can be thresholded. The thresholding process will produce a binary
image comprising two classes, namely road and non-road. These binary images typically
contain a number of artifacts which can be reduced through filtering.
In chapter 3, we have been introduced with mathematical morphology and several basic
filtering operators. The following two sections discuss two additional operations used by
the road network extraction system proposed in the subsequent chapters.
4.5.1 Connected Component Filter
The concept of connected components is rooted in graph theory and is defined as a sub-
graph within an undirected graph in which all vertices are connected with a path. In image
32
processing, a connected component is a set of pixels where each pixel has at least one
immediate neighboring pixel with the same value. The purpose of a connected component
filter is to remove or suppress objects larger or smaller than a specified threshold.
Connected components labeling scans an image and groups its pixels into components
based on pixel connectivity, i.e. all pixels in a connected component share similar pixel
intensity values and are in some way connected with each other. Once all groups have
been determined, each pixel is labeled with a graylevel or a color (color labeling)
according to the component it was assigned to.
Extracting and labeling of various disjoint and connected components in an image is
central to many automated image analysis applications.
Connected component labeling works by scanning an image, pixel-by-pixel (from top to
bottom and left to right) in order to identify connected pixel regions, i.e. regions of
adjacent pixels which share the same set of intensity values V. (For a binary image V={1};
however, in a graylevel image V will take on a range of values, for example: V={51, 52,
53, ..., 77, 78, 79, 80}.)
Connected component labeling works on binary or graylevel images and different
measures of connectivity are possible. However, for the following we assume binary input
images and 8-connectivity. The connected components labeling operator scans the image
by moving along a row until it comes to a point p (where p denotes the pixel to be labeled
at any stage in the scanning process) for which V={1}. When this is true, it examines the
four neighbors of p which have already been encountered in the scan (i.e. the neighbors (i)
to the left of p, (ii) above it, and (iii and iv) the two upper diagonal terms). Based on this
information, the labeling of p occurs as follows:
If all four neighbors are 0, assign a new label to p, else
if only one neighbor has V={1}, assign its label to p, else
if more than one of the neighbors have V={1}, assign one of the labels to p and
make a note of the equivalences.
After completing the scan, the equivalent label pairs are sorted into equivalence classes
and a unique label is assigned to each class. As a final step, a second scan is made through
33
the image, during which each label is replaced by the label assigned to its equivalence
classes. For display, the labels might be different graylevels or colors.
4.5.2 Closing
The dilation and erosion operations, introduced in chapter 2, can be used in different
combinations to construct additional mathematical morphology operators. The closing
operator is an example of such combination and the closing of a given set A by a kernel B
is obtained by a dilation of A by B and an erosion of the result with B. The closing
operations can be defined as
This operation can be used to fill small holes within larger components.
4.6 Data
The data were collected from the internet and Google earth map from different areas. The
high resolution satellite imagery was achieved with the following sensors:
Panchromatic: The sensor has a single black and white band with a ground resolution of
61 cm at nadir and a spectral range of 445-900 µm.
Multispectral: The sensor has a ground resolution of 2.4 m at nadir and has four bands,
namely blue (450-520 µm), green (520-600 µm), red (630-690 µm) and near infrared
(760-900 µm).
The final pan-sharpened product combines the spatial information from the panchromatic
band with the spectral information of the multispectral bands into a spectrally and spatially
enhanced color composite with a ground resolution of 61 cm. The composite can be
created by selecting various combinations of the multispectral bands and the data used in
this thesis was created with the red, green and blue bands.
The software packages and frameworks used to implement the road network extraction
system and its components are considered in the next section.
34
4.7 Software
The overall road network extraction system was implemented using MatLab with the
specified image processing features. The complete source code in Matlab for the automatic
road network extraction from the given high resolution satellite image and also the
implementations of some of the basic algorithms were obtained from the author. A
software package was developed while working on this thesis.
4.8 Test System Specification
All experiments were conducted on a single machine with the following specifications:
CPU: Intel ® Pentium ® Dual Core E2180 @ 2.00 GHz
Memory: 1 GB, DDR 2, 667 MHz
Hard Drive: 160 GB, 7200 rpm, SATA
Operating System: Windows XP Professional V2002 SP 3
Conclusion
This chapter discussed the overall components required for the road network extraction
system. The detailed step-by-step implementation procedure will be shown in the next
chapter. The most significant components thus found out are therefore include the
following: the edge detector, segmentation and classification, connected component filter,
mathematical morphology processing and so on. Here a brief ideology for the system
development has been achieved.
35
Chapter-5
[Implementation]
This chapter presents a road extraction system and the step-by-step procedures of the
overall implementation of the system.
Introduction
Human beings possess the ability to perceive relationships between image objects that are
not necessarily adjacent with relative case. This ability enables the detection of elongated
features, such as roads and rivers, in remotely sensed imagery. But with the help of an
intermediary human operator, extraction of road network is not just good enough process
to be convenient. We must employ an automatic extraction of road network system for this
purpose. This chapter presents a road extraction system and the step-by-step procedures of
the overall implementation of the system.
5.1 The System Pipeline
Our proposed approach comprises with two modules. In the first module, the road network
is established through segmentation, feature extraction and classification. Then in the
second module, the established road network is extorted through morphological processing
and linear regression.
Figure 5.1 shows the general process of road network extraction from high resolution
satellite image in sub-urban areas.
The input of the method is high resolution satellite image. The Module-1 first converts the
input true color image into grayscale image. After filtering the grayscale image using
smoothing filter (median filter), edge detection is performed to sense the edges of all
objects in the image. After edge detection, all of the edges are segmented to be connected
components through automatic segmentation. Then according to the filled area, qualified
connected components are selected to establish road network.
36
In the Module-2, the established road network image is filtered first to remove small
unnecessary objects. Next, different morphological operation, dilation and boundary
extraction are performed on the existing objects to eliminate the excess parts of image
objects and to grow sharp and full road network [43]. Finally linear regression is
performed in order to acquire accurate road image.
Figure 5.1: The pipeline to road network extraction
Input High Resolution Satellite Image
Grayscaling
Thresholding
Edge Detection
Convolution Filtering
Automatic Segmentation
Component Classification
Established Road Network
Smoothing Filtering
Morphological Processing
Output Road Network Image
Mod
ule-
1 M
odul
e-2
37
5.2 Input High Resolution Satellite Images
The input images were chosen randomly from the images available in the internet
produced by the Satellite. The input images taken for the implementation of the road
network extraction process are shown in the Figure 5.2.
Figure 5.2: Input high resolution satellite images
5.3Grayscaling
The input is true color image or RGB image. By Grayscaling system with the formula
Y = 0.3R+0.59G+0.11B, the RGB image can be gray scaled. The number of gray level in the
colormap is 256. The gray image is then be used as two-dimensional matrix and whose
elements denote the intensity values ranged from 0 to 255.Figure 2 shows some input true
color images and figure 3 shows the corresponding gray images.
38
Figure 5.3: Grayscale images
5.4 Thresholding
In order to automatically perform histogram shape-based image thresholding, Otsu’s
method can be used; that results in gray level image to a binary image [44].Figure 4
showsa binary image after performing thresholding. The algorithm assumes that the image
to be thresholded contains two classes of pixels or bi-modal histogram (e.g., foreground
and background). Then calculate the optimum threshold separating those two classes so
that their combined intra-class variance is minimal [45].
So, we exhaustively search for the threshold that minimizes the intra-class variance,
defined as a weighted sum of variances of the two classes:
훔훚ퟐ = 훚ퟏ(퐭)훔ퟏퟐ(퐭) + 훚ퟐ(퐭)훔ퟐퟐ(퐭)
Weights 훚퐢are the probabilities of the two classes separated by a threshold and 훔퐢ퟐ
variances of these classes.
Here’s the overall algorithm:
1. Compute histogram and probabilities of each intensity level
2. Set up initial훚퐢(ퟎ)and 훍퐢(ퟎ)
3. Step through all possible thresholds maximum intensity
1. Update 훚퐢and 훍퐢
2. Compute 훔퐛ퟐ(퐭)
39
4. Desired threshold corresponds to the maximum 훔퐛ퟐ(퐭)
Figure 5.4: Binary image
5.5 Edge Detection
For detecting the edges in an image, edge detection is performed. There are three
fundamental steps performed in edge detection:
1. Image smoothing for noise reduction.
2. Detection of edge points to extract all points that are potential candidates to
become edge points.
3. Edge localization to select from candidate edge points only the points that are true
members of the set of points comprising an edge.
Figure 5 shows the images after being edge is detected. For the purpose of finding
edges, edge detection is accomplished using first-or second-order derivatives [40].
5.6 Convolution Filtering
Convolution filtering is the linear filtering which is the mathematical foundation of
filtering. The process is:
1. Rotate the filter mask by 180 degree.
2. Move the mask over the image.
3. Compute the sum of products at each location
40
Figure 5.5: Image after edge detection.
The mask used here for filtering is called convolution kernel. In our approach, we used the
image to be filtered is used as convolution kernel as well. Doing this operation results in
reduced number of connected components.
5.7 Automatic Segmentation
Automatic segmentation is performed to segment objects from an image. To segment an
object, closed region boundaries are required. The desired edges are the boundaries of
such objects. From edge detection what we obtain are disconnected edges. In our
approach, disconnected edges are connected through segmentation and all the possible
objects of the image are broken into connected components. So, the result from the
automatic segmentation is the extraction of connected components of all the possible
objects in the image. Figure 5 shows some of the connected components of the image 5
shown in figure
41
.
Figure 5.6: Connected components
5.8 Component Classification
After performing segmentation, we have a lot of connected components of all the objects
in an image extracted. So, we need to classify only the connected components of the road
areas with other objects in the image. For example, the image shown in figure 4 has 54
connected components. But only 6 components are from road areas. So, in order to select
those 6 components among 54 components, using classification technique with the
comparison in some properties can be applied. Based on the filled area by the number of
on pixels, components of road can be selected. That means those components that have
larger filled area by the on pixels than a specified minimum value.
42
5.9 Established Road Network
After performing the component classification, all the eligible connected components of
road areas are integrated together to establish the road network. Figure 7 shows the
established road network for the image shown in figure 1 (left image) and it’s negative
image for achieving the goal.
5.10 Smoothing Filtering
Median filtering is a nonlinear process useful in reducing impulsive or salt-and-pepper
noise. It is also useful in preserving edges in an image while reducing random noise.
Impulsive or saltand pepper noise can occur due to a random bit error in a communication
channel.
Figure 5.7: Established road image and its complement image
In a median filter, a window slides along the image, and the median intensity value of the
pixels within the window becomes the output intensity of the pixel being processed. For
example, suppose the pixel values within a window are 5, 6, 55, 10 and 15, and the pixel
being processed has a value of 55. The output of the median filter in the current pixel
location is 10, which is the median of the five values. That means the original value of the
pixel is are included in the computation of the median [43].
43
Figure 5.8: Median filter image
5.11 Morphological Processing
Mathematical morphology is mainly set theory. Set reflection and translation are
employed extensively in morphology to formulate operations based on structuring
elements: small sets or sub-images used to probe an image under study for properties of
interest [40].
There are two Morphological operations Dilation and Boundary extortion are performed in
this step. The Dilation operations potentially filling in small holes and connecting disjoint
object. The dilation processes performed by laying the structuring element B on the mage
A. The structuring element can be square, rectangular, circular disc and any other shape
[46].
5.11.1 Dilation
The dilation operation on the image A by the structuring element B can be defined as:
, Where z is a displacement of the structuring
element. The value of the output pixel is the maximum value of all the pixels in the input
pixel's neighborhood. In a binary image, if any of the pixels is set to the value 1, the output
pixel is set to 1. Pixels beyond the image border are assigned the minimum value afforded
by the data type. For binary images, these pixels are assumed to be set to 0. For grayscale
images, the minimum value for uint8 images is 0. Dilation was performed to bridge the
gaps.
44
5.11.2 Erosion
The erosion operation on the image A by the structuring element B can be defined as
where z is a displacement of the structuring element. The
value of the output pixel is the minimum value of all the pixels in the input pixel's
neighborhood. In a binary image, if any of the pixels is set to 0, the output pixel is set to 0.
Pixels beyond the image border are assigned the maximum value afforded by the data type.
For binary images, these pixels are assumed to be set to 1. For grayscale images, the
maximum value for uint8 images is 255.
To dilate and erode, the input image to the dilation and erosion operation must be
processed first (grayscale, binary or packed binary image). These two basic morphological
operations can be combined in various ways to obtain other morphological operations like
opening, closing, and hit-or-miss transformation.
5.11.3 Closing
45
Fig 5.9: The resultant images after closing operation
In mathematical morphology, the closing of a set (binary image) A by a structuring
elementB is the erosion of the dilation of that set,
where and denote the dilation and erosion, respectively.In image processing, closing
is, together with opening, the basic workhorse of morphological noise removal. Opening
removes small objects, while closing removes small holes.
Thus, closing of A by B is the dilation of A by B, followed by the erosion of the result by
B (the same structuring element). The image obtained after closing is shown in the figure
5.9.
Conclusion
This chapter presents a road extraction system to test both pixel and object-based analysis.
The system was based primarily on the following approaches: the image segmentation
method, the classification system, the road network construction and the mathematical
morphology.
Though the automaticity has been achieved in this system, but still there is some scopes to
grow the improvement in the system. Instead of using a single classifier, a higher
extraction quality can be achieved by fusing the results from multiple classifiers.
46
Chapter-6 [PERFORMANCE EVALUATION]
This chapter introduces a qualitative and quantitative evaluation method for the
accuracyassessment of the automatic road network extraction system. The effectiveness
was also measured by the comparison of the result with the input.
6.1 Evaluation Method
From an algorithmic point of view, the extraction accuracy typically defines the success of
a feature extraction method. Accuracy is commonly measured bycomparing the algorithm
output against the reference data either from the GIS database orthrough manually
digitizing. According toWiedemann Hu [47],three indexes to evaluate the quality of road
extraction are as follow:
Correctness =LmLr(6.1)
Completeness = LmLr(6.2)
Quality = Lm( Lur+Le ) (6.3)
where Lr is the number of pixels composing of the reference road R, Le is the number of
thepixels composing of the extracted road V, Lmeis the number of thepixels composing the
extracted road that is in the reference buffer, which is the area in which the distance
between pixels and the reference roadis less than a given tolerance T. Lmr is the number of
the pixelscomposing the reference road that is in the extraction buffer, which is the area in
which the distance between pixels and extracted road is less than agiven tolerance T. Lm =
min (Lme, Lmr), and Lur is the number of the reference roads that areout of the extraction
buffer. In this study the width of the road is selected as the tolerance T.
6.2 Evaluation Mechanism
For the purpose of measurement of the performance, accuracy and effectiveness, a special
technique was implemented and that is superimposing. Superimposing means imposing
something ( hereby a image ) over something else as background. The system which was
implemented by us, accomplishes a similar effectiveness measurement mechanism for the
extraction of road network.
47
Finally after obtaining the output road network image, with our proposed approach we can
measure the accuracy of the extraction. That means how exact is the output road network
in terms of the input high resolution satellite image and whether the road lines are
extracted in the correct positions as in the input image. We can accomplish this by
superimposing the output road image over the input high resolution satellite image.
In this superimposing technique, the input high resolution satellite image is used as the
background image and the output road network image is used as the object image. The
superimposing will be totally transparent i.e., the background input image will still be
visible under the output object image. Then we can calculate the number of matched pints
between the input high resolution satellite image and the output road network image. In
our proposed approach, we have experienced that almost 90%-95% perfectness of the
extraction can be achievable. Table 7.1 shows the elapsed time and accuracy with respect
to the number of connected components in the high resolution satellite images.
Table 7.1: Experimental results of road images in terms of connected components
Image Location Total No. of Connected
Components
Elapsed Time
(Sec.)
Accuracy (%)
Rural 40-55 0.8-0.89 97-98
Highway 75-85 1.8-1.15 95-97
Sub-urban 120-150 1.55-1.7 92-95
Urban 280-350 2.5-3 85-90
48
Chapter-7 [CONCLUSION & RECOMMENDATION]
This chapter presents conclusions based on the findings of this study, and recommendations
for further research.
7.1 Conclusions
Our proposed approach comprises with two modules. In the first module, the road network
is established through segmentation, feature extraction and classification. Then in the
second module, the established road network is extorted through morphological processing
and linear regression.
The input of the method is high resolution satellite image. The Module-1 first converts the
input true color image into grayscale image. After filtering the grayscale image using
smoothing filter (median filter), edge detection is performed to sense the edges of all
objects in the image. After edge detection, all of the edges are segmented to be connected
components through automatic segmentation. Then according to the filled area, qualified
connected components are selected to establish road network.
In the Module-2, the established road network image is filtered first to remove small
unnecessary objects. Next, different morphological operation, dilation and boundary
extraction are performed on the existing objects to eliminate the excess parts of image
objects and to grow sharp and full road network. Finally linear regression is performed in
order to acquire accurate road image. A modified, fast and highly flexible method for the extraction of road network from very
high resolution satellite images has been presented. This method is composed of two steps.
In the first step, road network is established using the extracted connected components of
road areas and in the second step, the established road lines are rectified and further grown
up using mathematical morphology.
The main contribution of this work is to provide a modified method to automatically
extract road network from high resolution satellite images efficiently and flexibly. We
applied our proposed approach on seven different satellite images on urban, sub-urban and
49
rural area taken from internet and successfully extracted the overall road network to
determine the geo-graphical information.
The most attractive feature of our proposed approach is huge amount of flexibility (i.e.,
how easily and simply we can perform it) than any methods of earlier. The proposed
method has been successfully applied in the extraction of various types of road images.
7.2 Recommendations for Future Work
Further improvements can be made towards each individual component of the road
network extraction system. An interesting challenge will be to integrate more and more
automatic processes into the system.
In addition, morphological operations can be applied on a specific region, instead of
thewhole image. This helps the human operator to achieve better results.
Moreover, algorithms needs to be improved to be able to removed some branches which
arederived from the thinning process.
Again, the result of this road network extraction system can be easily inserted into GIS
database. The road information can be updated easily and then can also be viewed from
anywhere to identify the correct location.
Even sudden road crash can also be detected using this GIS updates containing the recent
road data.
Furthermore, the road network extraction system was developed by using MatLab which
has low computationefficiency. The implementation of the system using C++
programming may improvethat.
In summary, based on the automatic road extraction strategy, the system hasdemonstrated
to be efficient, accurate and reliable. With modifications suggested above, the automatic
road network extraction system may become of commercial value.
50
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
K. Lee and H. Ryu. Automatic circuitry and accessibility extraction by road graph
network and its application with high-resolution satellite imagery. In Proceedings of
the International Geoscience and Remote Sensing Symposium, 2004.
J. Sherman, J. Preisser, W. gesler, and T. Arcury. A suite of methods for
representing activity space in a healthcare accessibility study. International Journal
of health Geographics, 4:24, 2005.
J. Rocha and M. Queluz International of census data, remote sensing and GIS
techniques for land-use and cover classification. In Proceedings of SPIE – The
International Society for Optical Engineering, 2002.
B. Ayalew, R. Gomez, and W. R. O. Carrasco. Pavement management using
hyperspectral imagery. In Proceedings of SPIE- The International Society for
Optical Engineering, 2003.
Q. Zhang and I. Couloigner. A framework for road change detection and map
updating. The International Archives of the Photogrammetry, Remote Sensing and
Spatial Information Sciences, 34:729-734, 2004.
J. Sherman, J. Spencer, J. Preisser, W. Gesler, and T. Arcury. A suite of methods for
representing activity space in a healthcare accessibility study. International Journal
of Health Geographics, 4:24, 2005
Mena, J. B., 2003. State of the art on automatic road extraction for GIS update: a
novel classification. Pattern Recognition Letters, 24(16): 3037–3058.
Quackenbush, L.J., 2004. A review of techniques for extracting linear features from
imagery.Photogrammetric Engineering & Remote Sensing, 70(12): 1383–1392.
Zlatanova, S. and Li, J. (eds.), 2008. Geospatial Information Technology for
Emergency Response, ISPRS Book Series, Vol. 6, Taylor & Francis, London, 382pp.
Gorin, B., 2005. Performance metrics for pan sharpening methods and comparison
with ground truth, ASPRS Annual Conference, March 7-11, Baltimore, Maryland.
Vosselman, G., De Knecht, J., 1995. Road tracking by profile matching and Kalman
filtering. In: Gruen, Kuebler, Agouris, (Eds.), Workshop on Automatic Extraction of
Man-MadeObjects from Aerial and Space Images. Birkauser, Basel, pp. 265–274.
Niu, X., 2006. A semi-automatic framework for highway extraction and vehicle
detection based on a geometric deformable model. ISPRS Journal of
51
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
Photogrammetry and RemoteSensing 61: 170–186.
R. Ketting and D. landgrebe. Classification of multispectral image data by extraction
and classification of homogeneous objects. IEEE Transactions on Geoscience
Electronis, 14:19-26, 1976.
C. J. Robinove. Earth resources satellites and integrated resource surveys. In
Proceedings of the Symposium on Remote Sensing, pages 47-56, Pretoria, South
Africa, 1973.
Trinder, J.C. and Wang, Y., 1998. Automatic road extraction from aerial images.
Digital Signal Processing 8 (4): 215–224.
Mena, J. B., Malpica, J. A., 2003. Color image segmentation using the Dempster
Shafer theory of evidence for the fusion of texture. Pattern Recognition. Internat.
Arch.Photogrammet. Remote Sensing 34, Part 3/W8.
Chen, K.M., and Chen, S.Y., 2002. Colour texture segmentation using feature
distributions.Pattern Recognition Letters, 23: 755-771.
Asano, T., Chen, D.Z., Katoh, N., and Tokuyama, T., 1996. Polynomial-time
solutions to image segmentation. www. Portal.acm.org.
Definiens, 2008. DefinienseCognition homepage, URL: http://www.definiens.com/
(last date accessed: 22 April 2009).
Shneier, M. O., 1982. Extracting linear features from images using pyramids, IEEE
Transactions on Systems, Man and Cybernetics, July–Aug: 569–572.
Couloigner, I., and Ranchin, T., 2000. Mapping of urban areas: A multiresolution
modeling approach for semi-automatic extraction of streets, Photogrammetric
Engineering &Remote Sensing, 66(7): 867–874.
Bonnefon, R., Dhérété, P., and Desachy, J., 2002. Geographic information system
updating using remote sensing images, Pattern Recognition Letters, 23(9):1073–
1083.
Baumgartner, A., Steger, C., Mayer, H., Eckstein, W., and Ebner, H., 1999.
Automatic road extraction based on multi-scale, grouping, and context,
Photogrammetric Engineering &Remote Sensing, 65(7): 777-785.
Pigeon, L., Moulin, B., Solaiman, B., Toutin, T., and Thomdaughter, K. P. B., 1999.
Humanexperts features extraction strategies for topographical map production and
updating, Proceedings of International Geoscience and Remote Sensing Symposium,
28 June–02 July, Hamburg, Germany, pp. 296–298.
52
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
Mohammadzadeh, A., Tavakoli, A., and Valadanzoej, M., 2006. Road extraction
based on fuzzy logic and mathematical morphology from pansharpened IKONOS
images, ThePhotogrammetric Record, 21(113): 44–60.
Mokhtarzdae, M., and ValadanZoej, M.J., 2007. Road detection from high
resolutionsatellite images using artificial neural networks. International Journal of
Applied EarthObservation and Geoinformation, 9(1): 32-40.
Mckeown, D. M. and Denlinger, J. L., 1988. Cooperative methods for road tracking
in aerial imagery, in Proceedings of the IEEE Conference in Computer Vision and
PatternRecognition, pp. 662–672.
Shukla,V., Chandrakanth, R., and Ramachandran, R., 2002. Semi-automatic road
extraction algorithm for high resolution images using path following approach, in
ICVGIP02, Ahmadabad, Vol. 6, pp. 31-236.
Zhao, H., Kumagai, J., Nakagawa, M., and Shibasaki, R., 2002, Semi-automatic road
extraction from high resolution satellite images, ISPRS Photogrammet. Comput.
Vision, Graz, Ustrailia, September, pp. A-406.
Hu, J., Razdan, A., and Femiani J.C., 2007. Road network extraction and intersection
detection from aerial images by tracking road footprints, IEEE Transaction on
Geoscienceand Remote Sensing, 45(12): 4144-4157.
Kim, T., Park, S. R., Kim, M. G., Jeong, S. and Kim, K. O., 2004. “Tracking road
centerlines from high resolution remote sensing images by least squares correlation
matching,”Photogrammetric Engineering and Remote Sensing, 70(12): 1417-1422.
Zhou, J., Bischof, W. F., and Caelli, T., 2006 Road tracking in aerial images based
on human-computer interaction and Bayesian filtering, ISPRS Journal of
Photogrammetry &Remote Sensing, vol. 61: 108-124.
Kass, M., Witkin, A., Terzopoulos, D., 1987. Snakes: Active contour models.
Internat. J. Comput. Vis., 321–331.
Neuenschwander, W., Fua, P., Iverdaughter, L., Szekely, G., and Kubler, O., 1997.
Ziplock snakes. International Journal of Computer Vision 25 (3): 191–201.
Gruen, A. and Li, H., 1997. Semi-automatic linear feature extraction by dynamic
programming and LSB-snakes. Photogrammetric Engineering and Remote Sensing,
63 (8):985–995.
Trinder, J. and Li, H., 1995. Semi-automatic feature extraction by snakes, Automatic
Extraction of Man-Made Objects from Aerial and Space Images, BirkhaeuserVerlag,
53
[37]
[38]
[39]
[40]
[41]
[42]
[43]
[44]
[45]
[46]
[47]
pp.95-104.
Mayer, H., Laptev, I., Baumgartner, A., 1998. Multi-scale and snakes for automatic
road extraction. In: 5th Europ. Conf. on Comput. Vision, I: 720–733.
Laptev, I., Mayer, H., Lindeberg, T., Eckstein, W., Steger, C., and Baumgartner, A.,
2000. Automatic extraction of roads from aerial images based on scale space and
snakes. Machine Vision Applicat. 12 (1), 23–31.
Amo, M., Martinez, F., and Torre, M., 2006. Road extraction from aerial images
using a region competition algorithm. IEEE Transactions on Image Processing,
15(5):1192-1201.
R. Gonzalez, and R. Woods, Digital Image Processing, Addison-Wesley, New
York, 1993.
L. G. Sapiro and G.C Stockman. Computer Vision. New jersey, Prentice-hall, 2001.
PCI, 2004. Pan Sharpening in Geomatica 9, available at
http://www.Pcigeomatics.com/products/viewlets/pansharpening/pansharp_viewlet_fi
nal.swf.
Boshir Ahmed, Md. Ali Hossain and Md. Ariful Islam “Automatic extractions of
road intersections from satellite imagery in urban areas,” 6th International
Conference on Electrical and Computer Engineering (ICECE 2010), pp: 686-689,
2010 .
M. Sezgin and B. Sankur (2004). "Survey over image thresholding techniques and
quantitative performance evaluation". Journal of Electronic Imaging13 (1): 146–165.
Nobuyuki Otsu (1979). "A threshold selection method from gray-level histograms”.
IEEE Trans. Sys., Man., Cyber.9 (1): 62–66.
T. M. Talal1 , M. I. Dessouky2 , A. El-Sayed2 , M. Hebaishy1 & F. E. Abd El-
Samie2 “Road Extraction from High Resolution Satellite Images by Morphological
Direction Filtering and Length Filtering”, 1National Authority of Remote Sensing
and Space Science, Cairo (Egypt). 2Faculty of Electronic Engineering, Menoufia
University, Menouf, Egypt.
Wiedemmann, C., Heipke, C., and Mayer, H., 1998. Empirical evaluation of
automatically extracted road axes, in Proc. CVPR Workshop Empirical Eval.
Methods Comput. Vis., Los Alamitos, CA, pp. 172–187.